Welcome Note
This week’s issue comes to us from Ricardo Tavares, who advises governments and companies on technology policy and who wrote here in August on what was at stake in the U.S.-China AI talks. He returns to the industry he knows best. The AI build-out has taken the memory, and the bill is arriving on phones, laptops and games consoles, not only in the data centre. Ricardo has spent decades on the policy side of the semiconductor business, and his work will appear here again. Over to you, Ricardo.
A memory chip supplier has just asked the world to stop paying it so much. On September 9, Kioxia chief executive Hiroo Ota was reported saying that NAND flash “prices have already risen enough,” and that he had instructed his sales teams not to push for substantially higher prices from data center customers because “even hyperscalers have limited budgets.” This is the same Kioxia whose average NAND selling price rose 70% quarter-over-quarter in the three months to June, after more than doubling in the previous quarter.
The irony is hard to miss, and it is not the first of its kind. In July, at a business forum on Jeju island, SK Group chairman Chey Tae-won, whose SK hynix is the world’s largest maker of the high-bandwidth memory (HBM) feeding AI accelerators, said that memory prices were “abnormally high” and should fall, and that suppliers should not restrict supply to keep them elevated. A month later, he warned that “next year will be the most severe year for memory supply shortages,” that customers were requesting nearly double this year’s volumes, and, of the resulting inflation, that “this raises prices across society as a whole. It’s not a good thing.”
Why would the sellers of the scarcest commodity in the technology industry plead for restraint? The memory shortage has stopped being a story about data centers and has become a story about the price of a phone, a laptop, and a games console. Chey has warned that high memory prices are spilling into consumer prices; Ota has warned that further increases could damage long-term AI investment. How did the shortage come to be, who is paying for it, and why is one company, Apple, paying noticeably less than everyone else?
Enough Is Enough
The scale of the repricing has no precedent. The World Semiconductor Trade Statistics (WSTS) organization, the industry’s official statistics body, raised its 2026 memory revenue forecast at mid-year from $295 billion to $804 billion, a 2.7-fold revision, while barely touching its forecast for logic chips. Memory has gone from roughly 29% of all semiconductor revenue in 2025 to a forecast 53% in 2026. Those revenue forecasts claim that price, not volume, explains the rising share of memory in the industry sales.
The quarterly numbers are the ones buyers feel. Contract prices for conventional dynamic random-access memory (DRAM), the working memory in every PC, phone and server, rose 93% to 98% in the first quarter of 2026 alone, according to TrendForce. TrendForce’s reported first-quarter increase and its forecasts for the second and third quarters imply roughly a 3.5-fold increase in nine months. NAND flash, the storage memory in solid-state drives and phones, followed a quarter or two behind, which is where Kioxia’s 70% comes from.
Every major supplier has declared itself sold out. SK hynix said on October 29, 2025 that it had secured full customer demand for its 2026 DRAM and NAND production. In December, Micron said that, in the medium term, it expected to meet only “half to two-thirds” of demand from several key customers. Reports in August indicated 2027 output was already largely booked.
That is the context for Ota’s plea. Kioxia is not being generous. It is reading the demand side of its own order book, and what it sees there is buyers reaching what TrendForce calls their “affordability limit.”
The HBM Tax
The root cause sits inside the AI accelerator. High-bandwidth memory (HBM) is a stack of eight to sixteen thinned DRAM dies, bonded on top of a logic die and sealed into the same package as the graphics processing unit (GPU) it serves. It is the memory of choice for the chips made by Nvidia, AMD and the hyperscalers because it moves data over thousands of very short wires rather than a few long ones, which is far cheaper in energy per bit at data center scale.
The critical fact is that HBM is not a different product from a different factory. The DRAM in an AI accelerator and the DRAM in a $150 Android phone are manufactured in the same fabs and compete for the same wafer starts (the release of a new, blank semiconductor wafer into a fabrication line to begin the manufacturing process). A supplier that decides to make one more HBM stack is deciding to make several gigabytes fewer of everything else.
The exchange rate is brutal. Micron’s chairman and chief executive Sanjay Mehrotra told investors on December 17, 2025 that “the dramatic increase in HBM demand is further challenging the supply environment due to the 3-to-1 trade ratio with DDR5,” meaning that a gigabyte of HBM3E costs the fab roughly three gigabytes of conventional DRAM it could have produced instead. For the newer HBM4 generation, analysts put the ratio at four to one, a figure that should be read as an analyst estimate rather than any company disclosure. Two penalties explain it: an HBM die is physically larger per gigabyte stored, because a great deal of its silicon carries wiring and vias rather than memory cells, and a single failed stack destroys twelve or sixteen good dies at once.
The arithmetic follows. In 2026, HBM is estimated to absorb about 22% of the industry’s DRAM wafer starts in order to deliver about 9% of its bits; by 2027 those figures are forecast at roughly 30% and 13%.
What makes this cycle different is that the usual relief valve is missing. Past shortages ended when under-utilized fabs ramped from 70% or 80% utilization to 95% within a couple of quarters. Today there is no idle capacity, although Micron expects industry DRAM bit shipments to grow in the low-to-mid-20% range and NAND bit shipments by about 20% this year.
A new cleanroom takes three to five years, equipment lead times have stretched to 12 to 24 months, and 2026 DRAM capital expenditure rose only 14%, most of it directed to HBM process upgrades rather than new floor space. Micron’s chief financial officer Mark Murphy compressed the supply-side thesis into four sentences on March 18, 2026: “There are low inventory levels. There is declining bits-per-wafer on node advances. HBM’s trade ratio is increasing. Any new capacity really needs to be greenfield, which is a physical constraint that takes a lot of time.”
There is one twist worth watching. By the first quarter of 2026, conventional DDR5 server memory was earning more per wafer than HBM3E, and Mehrotra confirmed it on the same March call: “Yes, it is correct that the margins for non-HBM today are higher than HBM margins.” The premium product is no longer the premium margin, which hints that, at some price, capacity could rotate back toward the memory the rest of us buy.
Consumers Pay the AI Bill
The bill has landed on the kitchen table. HP expects memory and storage to rise from roughly 15% to 18% of its PC bill of materials to about 35% during fiscal 2026; its chief financial officer Karen Parkhill said on February 24, 2026 that, “We have seen memory costs increase roughly 100 percent sequentially.” Dell, Lenovo, ASUS and Acer all raised prices between December and the first quarter. Apple raised Mac, iPad and Vision Pro prices on June 25, 2026, citing memory. Sony added $100 to the PlayStation 5 in April, Microsoft raised Xbox prices worldwide in August.
Smartphones are where the shortage reaches the most people, and the numbers are the worst on record. The International Data Corporation (IDC) now forecasts that 2026 shipments will fall 16.7% to one billion units, the steepest decline ever measured, while the average selling price rises 27.6% to $581. Fewer phones, at higher prices, generating more revenue: that is textbook inelastic pass-through. IDC’s Francisco Jeronimo put it plainly: “The memory tsunami that we warned about is now hitting the market in full, and consumers are starting to pay the AI bill.” The 173 million phones that shipped below $100 last year, IDC says, “are facing an existential crisis,” and vendors are downgrading specifications from 12GB/256GB to 8GB/128GB in an attempt to hold price points.
The people who sell the chips inside those phones say the same thing. Qualcomm’s president and chief executive Cristiano Amon, on the company’s February 4, 2026 earnings call: “It’s 100% related to memory.”
The Chinese Android makers are the most exposed. In an April Weibo post, Xiaomi president Lu Weibing said that memory for a 12GB/512GB configuration cost the company roughly CN¥1,500 (about $220) more than in early 2025. OnePlus China’s president Li Jie put “this round of memory cost increases” above 400%, and warned that “for some time to come, overall prices will only get higher and higher.”
The sharpest single data point is inside one company. Samsung’s handset division swung to an operating loss of ₩0.7 trillion in the second quarter of 2026 on component costs, in the same quarter that Samsung’s memory division posted record profits. Samsung Mobile’s Daniel Araujo told analysts on July 29 that “the massive demand for AI servers is creating a shortage and driving up prices for mobile memory.” In this case, the company selling the chips and the company buying them are fortunate to share a headquarters.
The Apple Exception
Apple is not immune, but it is visibly insulated. Counterpoint Research’s second-quarter tally shows global shipments down 11% year over year, the lowest second quarter since 2013, with Xiaomi, OPPO and vivo all posting double-digit declines. Apple grew 3% and reached a record 20% share, having, by Counterpoint’s account, avoided price increases during the quarter. IDC’s full-year forecast is starker: Android shipments down 24.3% in 2026, iOS down 1.3%, with iOS gaining nearly four points of share to a record 23.6%.
The iPhone 18 Pro and Pro Max announced on September 9 are the test case, and Apple passed it more cheaply than expected. MacRumors translated TrendForce’s 14% to 18% price-rise forecast into a starting price of $1,249 to $1,299 for the iPhone 18 Pro, up $150 to $200, and reported TrendForce’s estimate said that the overall bill of materials for a 256GB Pro model had risen about 38%. Apple announced $1,199 for the Pro and $1,299 for the Pro Max. Apple’s $100 increase was below that estimated $150 to $200 range, although its undisclosed component costs make the effect on margins impossible to quantify. A $100 increase on a $1,099 phone is about 9%. Xiaomi’s CN¥1,500 (~$224) memory increase on a single mid-range handset is three-quarters of the CN¥2,000 (~$299) that Lu says an entry-level phone must now sell for.
Four things explain the difference.
The first is scale in the negotiating room. For instance, Apple is Kioxia’s single largest customer at roughly 20% of its revenue, which is why the headline read Ota’s restraint as an “implied reprieve” for Apple even though Ota never named it. Two days earlier, TrendForce carried a report that Apple had signed a NAND long-term agreement (LTA), with Kioxia identified by market watchers as the likely counterparty. The counterparty, duration, volume, pricing and allocation terms remain undisclosed.
The second is Apple’s willingness to consider Chinese memory both as an alternative source for devices sold in China and as leverage in negotiations. In February, Commercial Times reported that Apple was evaluating Chinese memory from CXMT and YMTC for the iPhone 18 series and MacBooks to “strengthen its negotiating position,” after Samsung sought a 100% price increase on iPhone LPDDR5X and SK hynix nearly the same. Apple has reportedly gone further, discretely lobbying in Washington to keep CXMT, now about 8% of global DRAM supply, off the U.S. Commerce Department’s Entity List despite congressional pressure. That is a reputational risk in Apple’s largest market, and its willingness to run it tells you how much the leverage is worth.
The third is the shape of the product line. Apple’s cheapest iPhone starts at roughly $600 and its average selling price sits well above the industry’s $581. A memory bill that rises by $200 is survivable inside a $1,199 device with a high gross margin and a customer base whose loyalty Counterpoint cites as a factor. The same $200 is existential for a $100 phone, which is why the sub-$100 segment is being abandoned rather than repriced. The shortage is regressive by design: it falls hardest on the products and the buyers with the least room. This is a disaster for digital inclusion.
The fourth is discipline. Apple holds flagship memory at 12GB while Android flagships moved to 16GB, and it concentrates scarce allocation on current-generation products rather than a wide catalogue. Samsung, the one Android maker with comparable scale, reclaimed the number one position in the second quarter through vertical integration and fewer price rises, but its handset division lost money while its memory division set records. Every other Android maker is negotiating quarterly, from a weaker position, for the memory that Apple’s contracts have already claimed.
Losing Memory
What the industry is losing is the conventional memory that every device other than an AI accelerator depends on, and the price of losing it is being paid by the people with the least bargaining power. From the third quarter of 2026, long-term agreements cap price exposure for the hyperscalers and the largest device makers, and TrendForce states plainly that “the primary source of server DRAM price increases will shift toward customers without LTAs.” That means smaller cloud providers, corporate IT departments, automakers and the Android makers below Samsung; it also means the buyer of a $100 phone, who will soon not be able to buy one.
Governments have noticed but have not acted. On June 3, 2026, nine industry associations representing automakers, retailers, broadband providers and medical device makers wrote to the U.S. Treasury and Commerce Departments to warn that “an urgent imbalance in the market for memory chips could lead to significant and sustained near-term price increases for American households.” On July 1, SEMI, whose members include all three major memory suppliers, pushed back, warning that “interventions that distort pricing or capacity decisions risk prolonging the demand downturn.” As of this writing there are no government stockpiles, allocation orders, or memory-specific emergency measures in any major jurisdiction. Governments are behaving as investors, not allocators, and the fabs they are subsidizing land only in 2028 at the earliest.
The consensus on the duration of the memory chip shortage is “tight through 2027, first real relief in 2028.” Micron says beyond calendar 2027, Samsung says worse in 2027, Intel’s chief executive Lip-Bu Tan says no relief until 2028. The forecasts of 2030 and beyond come from suppliers who are simultaneously asking customers to sign five-year prepaid contracts, a conflict of interest worth weighing. And even when the physical turn arrives, prices typically follow it by 12 to 18 months because of annual contract cycles.
Which brings us back to Ota and Chey. Two executives whose companies are printing money have concluded, apparently independently, that the money is coming from the wrong place. Chey has explicitly warned that high memory prices are spilling into consumer prices; Ota has warned that they could damage long-term AI investment. The demand being destroyed is the consumer’s, and the consumer has no contract. A supplier who understands that can see a 2028 in which the fabs finally open and the customers have gone. “Prices have already risen enough” is not generosity. It is the first sign that the memory makers have noticed who is paying the AI bill, and that they would rather not be holding it when it comes due.
We would love to hear your thoughts!
Ricardo Tavares
Guest Contributor, The Compute
Ricardo Tavares advises on public policy and regulation in AI, semiconductors and communications technologies. Since 2004 his work has concentrated on U.S., EMEA and Asia. He has negotiated with governments on behalf of Google, Ericsson and ByteDance, designed the privatisation of state-owned telecoms operators for the World Bank, and advised Inmarsat, GSMA, Vodafone, Telenor Group and Confidential Core AI, the latter a startup working on security for AI agents.
A note to our readers
Tokenando is an official media partner of AI Summit Barcelona, and we will be covering it in The Compute. The second edition runs on September 22 and 23, 2026 at the World Trade Center Barcelona.
As part of AI Week, we will be organising a workshop on the economics and regulatory aspects of AI in collaboration with our friends at BeLaw Lab and RiskOnyx
Follow the Token: from budgeting to legal compliance will take place on Monday, September 21, 2026 from 16:00 to 18:00 CEST at our offices in Cloudworks Passeig de Gracia. You can sign up on this link: https://luma.com/1l4nb64u
On Wednesday, September 23, 2026 at 16:15 CEST on the Panel Stage, I will be moderating Panel 19: From BSC to Brussels: Europe’s Supercomputing Bet on AI.
Readers of The Compute get 10 per cent off tickets with the code THECOMPUTE10 at aisummitbarcelona.com.
Wirelessly yours,
Ziad Matar
Co-founder, Tokenando
Editor-in-chief, The Compute
References
Kioxia Ota: https://wccftech.com/apple-wins-an-implied-reprieve-as-kioxia-ceo-declares-nand-prices-have-risen-enough/ (Ota’s words reach us via a P Equity Research post on X, September 9, 2026, and a TechNews Taiwan write-up: https://technews.tw/2026/09/09/kioxia-ceo-declares-nand-prices-have-risen-enough).
Chey Tae-won, July 2026 (”abnormally high”): https://www.trendforce.com/news/2026/07/20/news-sk-group-chairman-says-abnormal-memory-prices-should-fall-ai-chip-demand-seen-up-60-100-next-year/ (paraphrased by TrendForce from ETNews and Maeil Business)
Chey Tae-won, Jeju forum, August 2026: https://finance.biggo.com/news/04e0233f-20f2-455c-884b-911a2191cec8 (translated from Korean; secondary)
IDC 2026 forecast, August 26, 2026: https://www.idc.com/resource-center/blog/smartphone-shipments-set-for-record-16-7-drop-in-2026-as-the-memory-crisis-hits-full-force/
Counterpoint Q2 2026: https://counterpointresearch.com/en/insights/global-smartphone-shipments-q2-2026
TrendForce iPhone 18 price estimate (via MacRumors, September 3, 2026): https://www.macrumors.com/2026/09/03/iphone-18-pro-prices-apple-event-next-week/
Apple September 9 event prices: https://www.macobserver.com/tips/round-ups/apple-event-2026-by-the-numbers-every-price-date-spec/ (second source: https://www.techradar.com/news/live/apple-event-september-2026-live-blog)
Apple NAND LTA with Kioxia (Economy Tribune via TrendForce, September 7, 2026): https://www.trendforce.com/news/2026/09/07/news-apple-said-to-have-signed-nand-lta-with-kioxia-seen-as-likely-partner-foldable-iphone-memory-suppliers-in-focus
Apple evaluating CXMT/YMTC (Commercial Times via TrendForce, February 26, 2026): https://www.trendforce.com/news/2026/02/26/news-apple-reportedly-eyes-chinese-memory-as-leverage-against-samsung-and-sk-hynix-amid-rising-costs/